On 14.06.2022 18:00, Anthony PERARD wrote:
> On Tue, Jun 14, 2022 at 05:40:27PM +0200, Jan Beulich wrote:
>> GNU make 3.82 apparently has a quirk where exporting an undefined
>> variable prevents its value from subsequently being updated. This
>> situation can arise due to our adding of -rR to MAKEFLAGS, which takes
>> effect also on make simply re-invoking itself. Once these flags are in
>> effect, CC (in particular) is empty (undefined), and would be defined
>> only via Config.mk including StdGNU.mk or alike. With the quirk, CC
>> remains empty, yet with an empty CC the compiler minimum version check
>> fails, breaking the build.
>>
>> Move the exporting of the various tool stack component variables past
>> where they gain their (final) values.
>>
>> Signed-off-by: Jan Beulich <[email protected]>
>> ---
>> There may be further factors playing into the described quirk, as I've
>> also observed that simply running make as 2nd time would lead to
>> successful building of xen/.
>>
>> While this wasn't a problem until several weeks back, I've not been able
>> to identify which exact commit would have caused the breakage. Hence no
>> Fixes: tag.
>
> Reviewed-by: Anthony PERARD <[email protected]>
Thanks.
> Looks like this happened before: be63d9d47f ("build: tweak variable exporting
> for make 3.82")
Ah yes. I did think I had to deal with that before, but I did check patches
only back to early 2021. But it's somewhat worse than described there: It's
not just the origin which changes, but (as explained) it actually prevents
the variable to further change its value.
> So, maybe the issue is started again with 15a0578ca4 ("build: shuffle
> main Makefile"), which move the include of Config.mk even later.
Yes, that's certainly it. Thanks for spotting - I'll add a Fixes: tag.
Jan
